Geometry-aware multimodal large language models (MLLMs) often lean on external 3D encoders to supply structural priors, but running those extra encoder paths at inference adds significant latency and compute. GeoThinker-F proposes an alternative: instead of calling an external expert, the model learns to generate its own predictive latent world representations from its internal hidden states. The authors argue that intermediate semantic representations already hold enough information to predict useful geometry and future semantics, without needing to exactly reproduce a dedicated 3D encoder's features.

The framework uses a decoupled two-stage training strategy. In the first stage, the MLLM backbone stays frozen while lightweight Latent Frame Prediction (LFP) heads are trained against an offline 3D geometry expert and next-frame semantic targets. In the second stage, the external expert is removed entirely, and the model is jointly optimized to integrate its self-predicted priors through a Spatial-Grounded Temporal (SGT) Fusion module. This lets the model internalize spatiotemporal perception rather than relying on an external path during deployment.

Across diverse spatial intelligence benchmarks, GeoThinker-F maintains competitive performance while eliminating the runtime cost of external 3D encoders. The reported inference speedup is up to 3.6x, with latency dropping from 1.45 seconds to 0.40 seconds. The work suggests that predictive internal representations can replace explicit external structural priors, offering a faster route to spatial reasoning in MLLMs.